PolSAR Image Classification Based On Deep Polarimetric Data Feature Learning And Multiple Feature Fusion

Polarimetric synthetic aperture radar(PolSAR)has richer data acquisition ability of multifrequency,multi-polarization,day-night,and all-weather.It offers more efficient and reliable means of information collection for the earth surface observation.The fine surface classification of PolSAR is the important part in the research of remote sensing image interpretation,which aims to assign different category labels to image pixel units according to their scattering characteristics.The research of effective classification algorithms can help to mine the discriminative information in PolSAR images and improve the ability to recognize targets,thereby promoting the wide application of PolSAR in practical engineering.The dissertation respectively takes single-frequency PolSAR(SF-PolSAR)and multifrequency PolSAR(MF-PolSAR)images as the research objects.Starting from two main directions of complex-domain polarimetric data feature extraction and multi-frequency complementary information mining and fusion,the PolSAR image classification algorithms based on deep polarimetric data feature learning and multiple feature fusion are studied.The main purpose is to fully exploit the discriminative feature representation to achieve accurate classification of SF-PolSAR and MF-PolSAR images.The main contents and innovations can be summarized as the following four parts:1.In view of the limited ability of existing complex-valued neural networks to mine deep features and inability to perform pixel-level classification,a pixel-level PolSAR image classification algorithm based on complex-valued full convolutional neural network(CVFCN)is proposed.Through the multi-hidden layer stacking and layer-by-layer processing,the proposed complex downsampling-then-upsampling structure realizes the learning and fusion of multi-level polarimetric data features,so as to improve the discrimination ability of features and achieve pixel-to-pixel training and prediction.In addition,the constructed complex max-uppooling layer can greatly capture context information to reduce the influence of inherent speckle.To implement fast and efficient training,a complex-valued weight initialization strategy is proposed to initialize CV-FCN.Finally,the complex average cross entropy(CACE)loss is used to achieve faster convergence and more refined target discrimination.Experimental results on measured SF-PolSAR images confirm that CV-FCN has certain advantages in regional category consistency and boundary preservation,and effectively improves the classification accuracy and prediction efficiency.2.The presence of speckles and the absence of discriminative features make it difficult for pixel-level PolSAR image classification to achieve more accurate and coherent interpretation results,especially in the case of limited available training samples.To this end,a composite kernel-based elastic net classifier(CK-ENC)is proposed.Based on superpixel segmentation of different scales,three types of features are firstly extracted to fully exploit intrinsic characteristics of PolSAR data and capture richer discriminative information.Then,a composite kernel(CK)is constructed to realize multi-kernel feature fusion in the same semantic space,thereby enhancing the discrimination and robustness capabilities of features.Finally,an elastic net classifier(ENC)integrated with CK(CK-ENC)is proposed to achieve better classification performance with limited training samples.Experimental results on measured datasets show that CK-ENC can achieve a balance between intra-class variation and inter-class interference,and the classification results have better regional connectivity while maintaining class details.In addition,the algorithm efficiently obtains reliable classification results even in the case of limited training samples.3.Aiming at the problem that existing MF-PolSAR classification methods insufficiently mine the complementarity among frequencies,the dual-frequency attention fusion network(DFAF-Net)is proposed.Based on the guidance of frequency attributes,the constructed frequency-aware attention block(FAB)module realize the learning of frequency-aware deep polarimetric data features,so as to fully mine the complementarity among frequencies.In addition,the adaptive feature fusion block(AFFB)module is proposed to adaptively fuse different frequency-aware features,which can effectively eliminate redundant information to obtain complementary information with stronger independence.The obtained fusion features are more compact within classes and separable between classes,thereby effectively improving the classification performance.Experiments on measured dual-frequency PolSAR datasets verify that DFAF-Net can effectively eliminate the inaccuracy of singlefrequency classification,and improve the accuracy and robustness of classifying various ground targets.4.The current deep learning methods for MF-PolSAR classification only consider local spatiality but ignore nonlocal spatial relationship.Therefore,the multi-frequency semantics and topology fusion network(MF-STFnet)is proposed.The cross-band interactive feature extraction module(CIFEM)is proposed to explicitly model the deep semantic correlation between bands.It extracts cross-band interactive features to realize the interactive fusion and enhancement of information between bands,thereby leveraging the inter-band complementarity to make ground objects more separable.In addition,the graph sample and aggregate network(GraphSAGE)is constructed to dynamically capture the representation of nonlocal topological relationship of data structures in different bands.Therefore,the robustness of classification can be further improved by exploiting nonlocal spatial information.Furthermore,an adaptive weighting fusion(AWF)strategy is proposed to fuse inference from different bands,thereby jointly making robust final multi-frequency joint classification decision.Experiment results on measured MF-PolSAR images show that MFSTFnet can effectively fuse the complementarity,and combine local and nonlocal spatial information to obtain more competitive classification performance than existing methods.